External Validation of a Brain Imaging Model for Predicting Psychosis in Clinical High-Risk Cohorts
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Importance
Structural magnetic resonance imaging (MRI) models for predicting psychosis in people at clinical high risk (CHR-P) are rarely tested in new cohorts, so whether and why they lose accuracy is unknown.
Objective
To test whether a structural MRI model predicts psychosis in external cohorts and adds to symptom ratings, whether its accuracy varies with participant characteristics, and what explains differences between cohorts.
Design
This prognostic study applied the PRONIA structural MRI model, re-estimated in its training sample (329 participants; 26 transitions), without refitting to 4 external cohorts. Data were analyzed from July 2025 to October 2026.
Setting
Early-recognition services in North America and Europe.
Participants
People at CHR-P in 4 external cohorts.
Exposure
Baseline MRI risk score.
Main Outcomes and Measures
Transition to psychosis by study-specific criteria (symptom duration, 1 week to 1 month); discrimination (area under the curve [AUC]), calibration, net benefit, added value beyond symptoms, and discrimination by participant characteristics, site, and outcome group.
Results
Among 685 external participants (mean [SD] age, 20.0 [5.1] years; 319 [46.6%] female; 106 transitions), the AUC was 0.53 in NAPLS-3, 0.56 in MUC-FRUE, 0.65 in ZInEP, and 0.69 in FePsy (pooled, 0.56; 95% CI, 0.51-0.62), compared with 0.74 in PRONIA. The calibration slope was 0.17 (95% CI, 0.01-0.34), no probabilities gave clear net benefit, and adding the MRI score to symptom ratings changed AUC by 0.01 (95% CI, −0.02 to 0.05). Only negative symptoms hinted at better discrimination, and not robustly ( q = .15). Participants without transition scored like those in PRONIA, whereas those with transition scored lower, so sensitivity fell from 0.92 to 0.57 with similar specificity (0.52 vs 0.55). Within-site comparisons, reweighting for age, sex, and symptoms, and alternative classifiers did not restore discrimination.
Conclusions and Relevance
Outside its training cohort, this MRI model predicted psychosis poorly, a loss confined to people who developed psychosis and unexplained by scanners or participant characteristics. People may reach psychosis by routes that differ between cohorts, only some of them visible in brain structure. Identifying those for whom an MRI risk score works, if any, requires far more transitions, counted by shared criteria.
Key Points
Question
Does a structural brain imaging model predicting psychosis keep its accuracy in other clinical high-risk cohorts, and if not, why?
Findings
In this prognostic study of 685 participants across 4 external cohorts, discrimination ranged from chance to moderate (area under the curve, 0.53-0.69 vs 0.74 in the training cohort), and the score added little to symptom ratings. Participants without transition scored as in the training cohort despite different scanners, whereas those with transition scored lower.
Meaning
Brain imaging may predict psychosis in only some people at risk, so it could help only if symptoms or other tests first identify these people.